A Low-Rank Approach to Off-the-Grid Sparse Superresolution

We propose a new solver for the sparse spikes superresolution problem over the space of Radon measures. A common approach to off-the-grid deconvolution considers semidefinite relaxations of the total variation (the total mass of the absolute value of the measure) minimization problem. The direct resolution of this semidefinite program (SDP) is, however, intractable for large scale settings, since the problem size grows as f(c)(2d), where f(c) is the cutoff frequency of the filter and d the ambient dimension. Our first contribution is a Fourier approximation scheme of the forward operator, making the TV-minimization problem expressible as an SDP. Our second contribution introduces a penalized formulation of this semidefinite lifting, which we prove to have low-rank solutions. Our last contribution is the FFW algorithm, a Fourier-based Frank-Wolfe scheme with nonconvex updates. FFW leverages both the low-rank and the Fourier structure of the problem, resulting in an O(f(c)(d) log f(c)) complexity per iteration. Numerical simulations are promising and show that the algorithm converges in exactly r steps, r being the number of Diracs composing the solution.